Defeat for chat GPT provider Open AI in Munich court

In a dispute over the use of song lyrics, the chat platform GPT operator OpenAI has suffered a defeat against the German collecting society GEMA . The Munich I Regional Court ruled in favor of the plaintiff GEMA, finding that the US corporation had infringed copyright on nine well-known songs.

The legal dispute specifically concerned the use of nine texts from some very well-known songs – including “Atemlos” by Helene Fischer, “Männer” by Herbert Grönemeyer, “Über den Wolken” by Reinhard Mey and “In der Weihnachtsbäckerei” by Rolf Zuckowski.

The texts were used to train Chat-GPT and were reproduced exactly, or at least very similarly, in response to simple queries to the system. The court considered this proof that the texts had been stored in OpenAI's systems. It ordered OpenAI, among other things, to cease storing the texts and using them in its models, to pay damages, and to disclose information about the usage and the revenue generated from it. The ruling is not yet legally binding.

It is considered likely that the ruling will be appealed and the case will go to further appeals. The final decision could have implications far beyond song lyrics, as expert Silke von Lewinski from the Max Planck Institute for Innovation and Competition explained before the ruling. She sees "fundamental importance for all works, be they literature, journalistic texts, music, visual art, photography, or any other works used for generative AI . The issue here is how existing laws should be interpreted."

Should the ruling be appealed and GEMA prevail in the final instance, this would shift the balance of power between the creative industries and technology companies somewhat in favor of authors and other rights holders, she asserted. "Before a text can be used for generative AI, the rights holders would then have to give their consent and would have the opportunity to receive compensation for it." This is likely precisely what GEMA aims to achieve with its lawsuit.

That the AI ​​was trained with the nine songs was undisputed in the process. What happened afterward, however, was a key question. Was the song data memorized—that is, saved and thus duplicated—or did the training with the data cause Chat-GPT to regenerate the lyrics without saving them?

The court took a clear stance, interpreting the fact that the system output the texts it had been trained on as proof that it must have memorized them. Random output was ruled out.